Registry for layers and their parameters/variables.
This represents the collection of all layers in the approximate Fisher information matrix to which a particular FisherBlock may belong. That is, we might have several layer collections for one TF graph (if we have multiple K-FAC optimizers being used, for example.)
class LayerCollection: Registry of information about layers and losses.
class LayerParametersDict: An OrderedDict where keys are Tensors or tuples of Tensors.
get_default_layer_collection(...): Get default LayerCollection.